1. Application of HMMs: Speech recognition “Noisy channel” model of speech

2. Speech feature extraction Acoustic wave form Sampled at 8KHz, quantized to 8-12 bits Spectrogram Time Frequency Amplitude Frame (10 ms or 80 samples) Feature vector ~39 dim.

3. Speech feature extraction Acoustic wave form Sampled at 8KHz, quantized to 8-12 bits Spectrogram Time Frequency Amplitude Frame (10 ms or 80 samples) Feature vector ~39 dim.

4. Phonetic model

5. Phones: speech sounds Phonemes: groups of speech sounds that have a unique meaning/function in a language (e.g., there are several different ways to pronounce “t”)

6. HMM models for phones HMM states in most speech recognition systems correspond to subphones –There are around 60 phones and as many as 60 3 context-dependent triphones

7. HMM models for words

8. Putting words together Given a sequence of acoustic features, how do we find the corresponding word sequence?

9. Decoding with the Viterbi algorithm

10. Limitations of Viterbi decoding Number of states may be too large –Beam search: at each time step, maintain a short list of the most probable words and only extend transitions from those words into the next time step Words with multiple pronunciation variants may get a smaller probability than incorrect words with fewer pronunciation paths Word model for “tomato”

11. Limitations of Viterbi decoding Number of states may be too large Beam search: at each time step, maintain a short list of the most probable words and only extend transitions from those words into the next time step Words with multiple pronunciation variants may get a smaller probability than incorrect words with fewer pronunciation paths –Use the forward algorithm instead of Viterbi algorithm The Markov assumption is too weak to capture the constraints of real language

12. Advanced techniques Multiple pass decoding –Let the Viterbi decoder return multiple candidate utterances and then re-rank them using a more sophisticated language model, e.g., n-gram model

13. Advanced techniques Multiple pass decoding –Let the Viterbi decoder return multiple candidate utterances and then re-rank them using a more sophisticated language model, e.g., n-gram model A* decoding –Build a search tree whose nodes are words and whose paths are possible utterances –Path cost is given by the likelihood of the acoustic features given the words inferred so far –Heuristic function estimates the best-scoring extension until the end of the utterance

14. Reference D. Jurafsky and J. Martin, “Speech and Language Processing,” 2 nd ed., Prentice Hall, 2008